Novel bearing fault diagnosis model integrated with dual-tree complex wavelet transform, permutation entropy and optimized FCM

Ping Ma, Hongli Zhang, Wenhui Fan, Cong WangJournal of Vibroengineering 2018 · DOI

Abstract: In order to enhance the capability of feature extraction and fault classification of bearings, this study proposes a feature extraction approach based on dual-tree complex wavelet transform (DTCWT) and permutation entropy (PE), using the fuzzy c means clustering (FCM) to identify fault types. The vibration signal of bearings can be decomposed into several wavelet components with DTCWT which can describe the local characteristics of vibration signals accurately. And the PE of each wavelet component, which can describe the complexity of a time series, is calculated to be regarded as the fault features. Then forming the standard clustering centers by the FCM, we defined a standard using the Hamming approach degree to evaluate the classification results in the FCM. In order to verify the effectiveness of the proposed approach, compared with two other typical signal analysis methods: ensemble empirical mode decomposition (EEMD) and variational mode decomposition (VMD), through extracting fault features, it required to identify the fault types and severities under variable operating conditions. The experimental results demonstrate that the proposed approach has a better accuracy and performance to diagnose a bearing fault under different fault severities and variable operating conditions. The proposed approach is suitable for a fault diagnosis due to its good ability to the feature extraction and fault classification.

Introduction

Bearings are the heart of rotating machinery, but diagnosing their health is notoriously difficult. Vibration signals from failing bearings are non-stationary and non-linear. Standard analysis tools often struggle when operating conditions change—for example, a feature that identifies a fault at 1700 RPM might disappear or look completely different at 1750 RPM.

Existing methods like Empirical Mode Decomposition (EMD) suffer from "mode mixing," while standard Discrete Wavelet Transforms (DWT) lack "shift invariance," meaning small timing shifts in the signal can lead to wildly different results. This paper introduces a robust bearing fault diagnosis model that combines Dual-Tree Complex Wavelet Transform (DTCWT) and Permutation Entropy (PE) with an optimized Fuzzy C-Means (FCM) clustering algorithm to create a pipeline that is robust to operational shifts yet sensitive to minute fault signatures.

Methodology

The proposed diagnostic pipeline consists of three primary stages:

  1. Decomposition (DTCWT): Unlike standard wavelets, DTCWT uses two parallel trees (real and imaginary) to achieve nearly shift-invariant signal decomposition. This ensures that the local characteristics of the vibration signal are captured accurately regardless of when the sensor starts recording.
  2. Feature Extraction (Permutation Entropy): The decomposed signals are converted into feature vectors using Permutation Entropy (PE). PE measures the complexity or "randomness" of a time series. A healthy bearing has a different complexity signature than one with a cracked inner race.
  3. Classification (Optimized FCM): The features are fed into a Fuzzy C-Means (FCM) clustering algorithm. The optimization introduces the Hamming approach degree (a measure of nearness). Instead of just assigning a sample to a cluster, the model calculates how closely a test sample matches the "standard" cluster center of a specific fault type.
Overview of the proposed fault diagnosis pipeline, illustrating the flow from raw vibration signals to DTCWT decomposition, PE feature extraction, and FCM clustering.

Experimental Setup and Datasets

The study utilizes the Case Western Reserve University (CWRU) Bearing Data Center dataset, a benchmark in the field. The hardware setup includes 6205-SKF deeply grooved ball bearings. The research evaluates the model across:

  • Fault Types: Inner race, rolling element, and outer race faults.
  • Severities: Fault diameters of 0.1778 mm, 0.3556 mm, and 0.5334 mm.
  • Operating Conditions: Motor loads ranging from 0 to 3 horsepower (speeds of 1720 to 1797 RPM).
The experimental motor setup (a) and sample vibration signals for various fault types (b) and severities (c).

Key Results and Discussion

The proposed approach demonstrated significant advantages over traditional methods like Ensemble Empirical Mode Decomposition (EEMD) and Variational Mode Decomposition (VMD):

  • 100% Accuracy: In tests identifying fault types under the same load, the proposed model achieved a 100% recognition rate.
  • Cross-Load Robustness: When trained on data from a 0 hp load and tested on data from 1 hp, 2 hp, and 3 hp loads, it maintained a 100% recognition rate. In comparison, EEMD and VMD methods misclassified several samples, often mistaking outer-race faults for inner-race faults.
  • Clustering Quality: Visual analysis of the feature space demonstrates that DTCWT-PE features form tight, distinct clusters that are easily separable, whereas other methods often see their clusters drift and merge as motor load increases.

Conclusion

The integration of DTCWT and PE provides a fundamental physical signature of bearing faults that is invariant to motor speed and load. By utilizing optimized FCM with Hamming nearness, the model achieves near-perfect classification accuracy. Future work will focus on optimizing the algorithm for real-time industrial applications where computational overhead is a primary concern.

Summary updated

Abstract

In order to enhance the capability of feature extraction and fault classification of bearings, this study proposes a feature extraction approach based on dual-tree complex wavelet transform (DTCWT) and permutation entropy (PE), using the fuzzy c means clustering (FCM) to identify fault types. The vibration signal of bearings can be decomposed into several wavelet components with DTCWT which can describe the local characteristics of vibration signals accurately. And the PE of each wavelet component, which can describe the complexity of a time series, is calculated to be regarded as the fault features. Then forming the standard clustering centers by the FCM, we defined a standard using the Hamming approach degree to evaluate the classification results in the FCM. In order to verify the effectiveness of the proposed approach, compared with two other typical signal analysis methods: ensemble empirical mode decomposition (EEMD) and variational mode decomposition (VMD), through extracting fault features, it required to identify the fault types and severities under variable operating conditions. The experimental results demonstrate that the proposed approach has a better accuracy and performance to diagnose a bearing fault under different fault severities and variable operating conditions. The proposed approach is suitable for a fault diagnosis due to its good ability to the feature extraction and fault classification.
BibTeX
@article{ma2018novel,
  author = {Ping Ma and Hongli Zhang and Wenhui Fan and Cong Wang},
  title = {{Novel bearing fault diagnosis model integrated with dual-tree complex wavelet transform, permutation entropy and optimized FCM}},
  journal = {Journal of Vibroengineering},
  year = {2018},
  doi = {10.21595/jve.2017.18278},
}
Figure